layer 3 low latency handoff

8/11/2019 Layer 3 Low Latency Handoff

http://slidepdf.com/reader/full/layer-3-low-latency-handoff 1/13

Mobile Networks and Applications 9, 633–645, 2004© 2004 Kluwer Academic Publishers. Manufactured in The Netherlands.

Performance Evaluation of Layer 3 Low Latency

Handoff Mechanisms ∗

C. BLONDIA∗∗University of Antwerp, Department of Mathematics and Computer Science, Middelheimlaan 1, B-2020 Antwerp, Belgium

O. CASALS and LL. CERDÀTechnical University of Catalonia (UPC), Department of Computer Architecture, Jordi Girona 1-3, D6, E-08034 Barcelona, Spain

N. VAN DEN WIJNGAERT and G. WILLEMSUniversity of Antwerp, Department of Mathematics and Computer Science, Middelheimlaan 1, B-2020 Antwerp, Belgium

Abstract. This paper investigates the performance of two Layer 3 low latency handoff mechanisms proposed by the IETF, namely Pre- Post-Registration. These protocols use Layer 2 triggers to reduce the built-in delay components of Mobile IP. We propose a simple analymodel that allows assessing the packet loss and the delay characteristics of these mechanisms. We describe several handoff implementatover a wireless access based on the IEEE 802.11 standard and analyze several implementation issues. Finally we study the scalability oprotocols using an OPNET simulation.Keywords: mobile IP, low latency handoff, IEEE 802.11, performance evaluation

1. Introduction

The growing number of portable computing devices and therequirement to provide seamless connectivity to the globalInternet using end-to-end IP solutions for mobile users havestimulated the research into IP mobility protocols. Mobile IP[19] is the current standard solution for mobility managementin IP networks. It allows a Mobile Node (MN) to change itspoint of attachment from one access router to another acrossmedia of similar or dissimilar types with the help of MobilityAgents. Establishment of new tunnels can introduce consid-erabledelays in the handoff processdue to the round-trip timebetween the Foreign Agent (FA) and the Home Agent (HA)during the registration process. Applied in an environmentwith frequent handoffs, this may lead to unacceptable distur-bance to ongoing sessions in terms of handoff latency andpacket loss.

The IETF has proposed a number of protocols that handlelocalmovements locally. These so-called micro-mobility pro-tocols incorporate a number of important design features re-lated to location management, routing and handoff schemes.They should fulll requirements such as simplicity to imple-ment, scalability with respect to the induced signaling, ef-∗This work was supported by the Ministry of Education of Spain under

grant CICYT TIC-2001-0956-C04-01 and by the Department of Univer-sities, Research and Information Society (Generalitat de Catalunya) un-der grant CIRIT 2001-SGR-00226, by the Fund for Scientic ResearchFlanders under project G.0315.01, by DWTC Belgium under project IAPP5/11 MOTION (Mobile Multimedia Communication Systems and Net-works) and by IWT under project 020152 – End-to-end QoS in IP basedMobile Networks.

∗∗Corresponding author.

ciency and performance with respect to packet loss and intro-duced delay.

Two approaches to hierarchical mobility managementhavebeen proposed [8]:

(i) Host-basedroutingschemes like Cellular IP (CIP) [6] andHAWAII [24] use mobile-specic routingwithout tunnel-ing to route packets inside the access network. We haveevaluated these mechanisms in [2,3] and [27].

(ii) Hierarchical tunneling techniques like Regional Tunnel-ing [13] rely on a tree-like structure of FAs. Trafc des-tined to the MN is encapsulated at the HA and then deliv-ered to the root FA. Each FA in the tree decapsulates andthen re-encapsulates the packets as they are forwardeddown the tree of FAs towards the MN. Location updatesat the adequate points of the tree follow the MN move-ments between different base stations so that trafc is

tunneled to the MN’s point of attachment.These protocols are designed without any assumption re-

garding underlying layers. This allows the widest possibleapplicability and a clean separation between Layer 2 andLayer 3 of the protocol stack. However, this layer separa-tion results in lower performance. Indeed, the MN may onlycommunicate with a directly connected FA and therefore itcan only start the registration process after completion of theL2 handoff. Moreover, the MN is unreachable during theregistration process, a property that may contribute to a non-negligible handoff latency and packet loss.

In this paper we analyze two mobility protocols [10] thataim at low latencyLayer 3 handoff based on Layer 2 informa-tion receivedby means of L2 triggers. In Pre-Registration, the



634 BLONDIA ET AL.

MN communicates with the new FA through the old FA forregistration purposes. In the second type, Post-Registration,the data is delivered to the MN by thenew FA, even before theactual registration process has been completed. These meth-ods can support both the normal Mobile IP model [19] in

which the MN is receiving packets from a HA and the Hi-erarchical Mobile IP model [12] in which the MN receivespackets from a Gateway Foreign Agent (GFA).

This paper focuses on the performance evaluation of theseschemes and their comparison. We describe possible imple-mentations of the low latency handoff mechanisms over anIEEE 802.11 network and analyze various implementation is-sues.

We present a simple analytical model that allows a detailedanalysis of the delay characteristics and the buffer require-ments for a single MN involved in a handoff. In particularwe investigate the inuence of the timing of the L2 triggers

on the delay and the packet loss of UDP streams towards anMN that is involved in a handoff. In order to investigate thescalability of the protocols, we have developed an OPNETsimulation model that allows considering more realistic sys-tems with respect to the number of MNs and their mobilitypattern.

2. Low latency handoff mechanisms: Pre- andPost-Registration

Mobile IP [19] denes how a MN performs a Layer 3 handoff between subnets served by different FAs. In order to keep aclear separation between L2 and L3 functionality and to al-low for the widest possible applicability, Mobile IP has beendesigned without any assumption concerning the underlyinglink layer. This separation between layers results in the fol-lowing sources of handoff delays. First, an MN involved in ahandoff may only begin the registration process after the L2handoff to the new Foreign Agent has been completed. Sec-ondly, as the messages generated by the registration processneed some time to propagate through the network, the MN isunable to send or receive packets during that time. This maylead to a handoff latency that is unacceptable for the supportof real-time services. For these reasons, the Mobile IP work-group of the IETF has proposed so-called low latencyhandoff schemes [10] based on information of the L2 handoff processreceived using L2 triggers [28]. Two types of schemes maybe distinguished, depending on the above mentioned prob-lem they address. A rst type allows the MN to register withthe new FA while still being connected with the old FA. ThePre-Registration scheme belongs to this class. In a secondtype, the packets can be delivered to the MN from the newFA before the registration process has completed. The Post-Registration scheme belongs to this class. In what follows,we give a description of both schemes. Details can be found

in [10]. First we introduce the reference architecture that isused throughout the paper.

Figure 1. Reference network.

2.1. The reference architecture

Consider the network conguration depicted in gure 1. Theaccess network is connected to the Internet via a gateway

router. The two access points (AP) belong to different subnetswith access routers referred to as respectively the old accessrouter and the new access router. The APs may or may notbe co-located with the access routers. The access routers andthe gateway router are provided with mobility agents, respec-tively the old Foreign Agent (oFA), the new Foreign Agent(nFA) and the Gateway Foreign Agent (GFA).

Remark that this reference network has no tree-like struc-ture as encountered in most networks. Router 3 has beenadded to allow independence between the distance (oFA,nFA) and the distance (nFA, GFA). This allows for signalingpackets exchanged between both FA to follow a shorter path.Note that this may also happen in a tree-like network. A moredetailed discussion on this issue, together with the impact of the network topology on the performance of the low latencyschemes, will be given in section 5.4.

We assume that a Corresponding Node (CN) sends pack-ets to the Mobile Node. We consider both Constant Bit Tate(CBR) UDP packet streams and TCP ows.

2.2. L2 triggers

The handoff schemes investigated in this paper make use of L2 triggers. Such a trigger is a signal related to theL2 handoff process. A rst trigger that is used is an early notice of an

upcoming change in the L2 point of attachment of the MN,referred to as anticipation trigger . A second trigger, the Link Down trigger (L2-LD), indicates that the L2 link between theMN and the old AP is lost. The Link Up trigger (L2-LU)occurs when the L2 link between the MN and the new AP isestablished. A trigger initiated at the old FA is referred as asource trigger and a trigger initiated at the new FA is referredas a target trigger .

For a detailed discussion on these L2 triggers, we referto [10].

2.3. The Pre-Registration handoff scheme

Pre-Registration realizes an anticipated L3 handoff. The net-work assists the MN in performing an L3 handoff before the



PERFORMANCE EVALUATION OF LAYER 3 LOW LATENCY HANDOFF MECHANISMS 635

Figure 2. Mobile-initiated Pre-Registration handoff protocol.

L2 handoff is completed. Both the MN (mobile-initiated) andthe FAs (network-initiated) can initiate a handoff.

A mobile-initiated handoff occurs when the L2 anticipa-tion trigger is received at the MN informing it that it willshortly move to the nFA. The L2 trigger contains informa-tion such as the nFA’s IP address (or a means to obtain it, e.g.,based on L2 addresses, see [10] for details).

The following messages are involved (see gure 2).

– Messages 0a and 0b contain a solicitation for a Router Ad-vertisement from the oFA to the nFA and a reply RouterAdvertisement from nFA to the oFA. The oFA should so-licit and cache advertisements from the nFA in advanceof the Pre-Registration handoff in order not to delay thehandoff. These messages will not be taken into account inthe performance analysis.

– Message 1a , a Proxy Router Solicitation, is issued by the

MN as a consequence of the L2 anticipation trigger.– Message 1b , a Proxy Router Advertisement, is sent by the

oFA as a result of the MN solicitation message.– Message 2 , the Registration Request to the nFA should be

sent via the oFA if the L2 handoff is not completed (i.e.,before L2-LD) or directly to the nFA if the L2 handoff isnished (i.e., after L2-LU).

– Message 3 , a Regional Registration Request with theGFA,issued by the nFA upon receiving GFA information withinmessage 2 from the MN.

– Message 4 , the Regional Registration Reply, sent by the

GFA to the nFA.Until the MN actually completes the L2 handoff to the new

FA and fully establishes the new L2 link, the nFA mayreceivepackets for the MN to which it does not have a direct linklayer connection. In that case the nFA can decide to drop orto buffer the packets for the MN.

Between the time that the Registration Reply is sent fromthe GFA to the nFA till the MN’s connection on L2 at thenFA is fully established, the GFA can also bicast trafc for theMN to both the oFA and the nFA (refer to [11] for a furtherdiscussion about this mechanism).

A network-initiated handoff can be initiated by a source

trigger at the oFA (source-initiated handoff) or by a targettrigger at the nFA (target-initiated handoff).

Figure 3. Post-Registration handoff.

A source-initiated handoff is initiated at the oFA by a re-ceived L2 trigger that informs the oFA of a MN’s upcomingmovement from oFA to nFA. The difference with the mobile-initiated case previously described is that Message 1a of g-ure 2 is not used.

A target-initiated handoff is initiated at the nFA by a re-ceived L2 trigger that informs the nFA of a MN’s upcomingmovement from oFA to nFA. The difference with the mobile-initiated case is that Message 1a of gure 2 is not used andthat Message 1b is sent by the nFA to the MN via the oFA.

2.4. The Post-Registration handoff scheme

The Post-Registration handoff method is based on a network-initiated model of a handoff which does not require any MNinvolvement until the actual L2 connection with the nFA iscompleted. The name of this technique nds its origin in thefact that the registration occurs after the L2 handoff has been

completed. This approach uses a bi-directional edge tunnel(BET) to perform a low latency change in the L2 point of attachment of the MN without requiring any involvement of it.

A handoff occurs when the MN moves from the oFA,where the MN performed a Mobile IP registration, to the nFA.The MN delays its registration with the nFA, while maintain-ing connectivity using the BET between the oFA and nFA.In [10], two different Post-Registration handoff schemes aredened: Sourceand Target TriggerPost-Registration. The se-quence of messages for both schemes is depicted in gure 3.

An FA becomes aware that a handoff is about to occur atL2 through the use of an L2 trigger. Two types of triggerscan be received: (i) a source trigger at the oFA (L2-ST) and(ii) a target trigger at the nFA (L2-TT). The FA receiving thetrigger sends a Handoff Request (HRqst) to the other FA. TheFA receiving the HRqst sends a Handoff Reply (HRply) to theother FA. This establishes a BET. The L2-LD (Link Down)trigger at the oFA and at the MN signals that the MN is notconnected anymore with the oFA. When the oFA receives theL2-LD trigger, it begins forwarding the MN packets throughthe forwarding tunnel to the nFA.

When the nFA receives the L2-LU (Link Up) trigger, itbegins delivering packets tunneled from the oFA to the MNand forwards packets from the MN. When the MN receivesthe L2-LU, it decides to initiate the Mobile IP Registrationprocess with the nFA by soliciting an Agent Advertisement



636 BLONDIA ET AL.

or continues using the BET. Once the Registration process iscomplete (through the exchange of a Regional RegistrationRequest and a Regional Registration Reply with the GFA),the nFA assumes the role of oFA.

3. Implementation of low latency handoff mechanismsover IEEE 802.11

We describe a possible implementation of the L3 handoff mechanisms described previously in a wireless network hav-ing an 802.11 link layer [1]. Our goal is to illustrate by meansof a realistic example the abstraction made in previous sec-tions about the L2 triggers and the interaction with the L3layer protocol. Other authors have also discussed the integra-tion of L3 and 802.11 handoffs, e.g., [16] and [12].

We rst introduce the 802.11signals issued duringhandoff and we describe how to interpret them as L2 triggers used to

implement various L3 handoff mechanisms. Subsequently,we discuss some implementation issues.

3.1. IEEE 802.11 handoffs

The IEEE 802.11 standard denes two operational modescommonly referred to as ad hoc and infrastructure . In ad hocmode MNs communicate on a peer-to-peer basis. This modeis usually conceived for a scenario where a group of MNs dy-namically set up a network to communicate among them. Ininfrastructure mode all MN transmissions go through an Ac-cess Point (AP). Usually, APs are not mobile and form part of a wired network.

We assume that the wireless network operates in infrastructure mode . In this mode, before an MN can send datapackets, it has to be authenticated and associated with an AP.The simplest authentication mechanismis the so-calledOpen-system authentication (the only method required by 802.11).This method is in fact a null authentication mechanism witha two-way message exchange between the MN and the AP.After being authenticated, the MN associates with the APby sending an Association Request frame (or Re-associationif the MN is already associated), which in turn is answeredwith an Association (or Re-association) Reply. In case of re-association, the Re-Association Request frame conveys theMAC address of the old AP. Another management frame usu-ally involved in 802.11 handoffs is the so-called beacon , sentperiodically by the APs (typically every 100 ms).

Handoffs are always initiated by the MNs and the stan-dard denes two procedures for the MN to decide which APto handoff: active scanning and passive scanning . Using ac-tive scanning, the MN sends Probe Requests frames on everyavailable channel. The APs answer these frames with ProbeResponses that are collected by the MN. Based on the proberesponses, the MN selects the AP that would make the bestconnection. Using passive scanning the MN listens for thebeacons sent by the APs in order to decide which one to hand-off. The event triggering the L2 handoff is implementationspecic (and may be initiated by the card rmware using pro-prietary algorithms).

Figure 4. Target trigger – Post-Registration handoff.

3.2. Pre- and Post-Registration over IEEE 802.11

The exibility of the handoff mechanism dened by the802.11 standard allows a large number of possible scenarios.For sake of simplicity, we shall assume the following: (i) TheAP is an embedded entity in the FA such that the L2 triggerscan easily be implemented by means of some internal inter-face. (ii) The MN uses the same channel with all APs. This al-lows for the following simple way to trigger the L2-handoffs:(iii) The movement at L2 layer is detected upon receiving therst beacon from the new AP. Note that a timeout is neededto avoid ping-ponging between both APs. (iv) We do not takeinto account the authentication phase. These assumptions arediscussed in section 3.3. Thetraces shown in this section havebeen obtained with the ns simulator [18].

3.2.1. Post-Registration handoff Figure 4 shows a trace with a possible handoff using the Post-Registration scheme. In this trace the CN sends a CBR streamto the MN. The gure shows: (i) the instants when the packetsare sent (indicated as “Tx by CN”); (ii) the instants when theMN receives the packets (these are indicated as “Rx (oFA)”and “Rx (nFA)” according to which FA they are sent by);(iii) the signaling packets sent during the handoff.

When the MN approaches the nFA, the L2 beacons sentby the nAP trigger the L2 handoff at the MN, which sendsa Re-association Request (RAReq in the gure) to the nAP.Upon receiving this frame, there is a target trigger at the nFA(LU trigger), which sends the Handoff Request (HRqst in thegure) to the oFA (in section 3.3 we explain a possible wayfor the nFA to obtain the IP address of oFA). Upon receivingthe HRqst, the oFA sends the Handoff Reply (HRply in thegure) and establishes a tunnel with the nFA. In this way, thepackets can reach the MN via the nFA after the coverage withthe oFA has been lost. The oFA starts sending the packetsaddressed to the MN through the tunnel immediately after itis established. This instant corresponds with the occurrenceof the LD trigger introduced in section 2.3. Finally, when

the nFA sends the Router Advertisement, the MN makes aRegistration with the nFA.




Figure 5. Target trigger – Pre-Registration handoff.

3.2.2. Pre-Registration handoff Figure 5 depicts a trace analogous to gure 4 for the Pre-Registration mechanism. The signaling packets sent duringthe handoff are the following. The new AP issues L2 beaconswhich are interpreted as L2 triggers at the MN indicating theimminent movement to this AP. Then the MN sends the ProxyRouter Solicitation (indicated as “MN PRsol”) to oFA. Uponreception of this packet, the oFA sends a Proxy Router Adver-tisement (indicated as “oFA PRadv”) to the MN. Upon recep-tion of this packet, the MN sends a Registration Request tothe oFA with destination address the nFA (indicated as “MNPRegReq”). Then, the MN sends the Re-association to thenAP.

A more detailed description of possible implementations

of Pre- and Post-Registration handoff mechanisms usingIEEE 802.11 can be found in [5] and [9].

3.3. Implementation issues

As mentioned before, the main goal of the implementationspreviously proposed was to clarify the abstraction of L2 trig-gers. Therefore, we have chosen simple assumptions in orderto keep our proposal as simple as possible. In fact, 802.11standard allows a wide variety of scenarios. Analyzing thepossible solutions for each of them goes beyond the scope of this paper. Anyhow, in the following we consider the appro-priateness of our assumptions given some scenarios that maybe found in practice.

First, we assumed that the AP is an embedded entity inthe FA. In some networks however, the FA and the AP maybe different entities. For example, a number of APs distrib-uted over a geographical area may be connected to the sameFA. In this case, a change from AP does not imply an L3handoff. This physical separation between the FA and the APwould prevent the Re-association Request sent by the MN tothe new AP to be the target trigger as described in the Post-Registration handoff. One possible solution could be the fol-lowing. After re-association an L2-source trigger takes placeat the MN, which in turn triggers the standard MIP registra-

tion process, i.e. the MN broadcasts an Agent Solicitation andsends a Registration Request to nFA uponreceiving the Agent

Advertisements from the candidate FAs. The MN should usethe PFANE (Previous Foreign Agent Notication Extension,see [22]) in the Registration Request. Knowing the Previ-ous Foreign Agent IP address, the nFA sends the Handoff Re-quest to the oFA. This mechanism is equivalent to the Smooth

Handoff described in [23].The second assumption, namely the MN uses the samechannel for all APs, may also not be satised in many sit-uations. This is because it is convenient to allocate adjacentAPs in non-interferingchannels, in order to increase themaxi-mum network load. In fact, a recent study [17]analyzed a real802.11 network where adjacent APs used different channels.MNs use active scanning to nd out new channels and avail-able APs when the 802.11 handoff is triggered. This studyshows that the scanning phase is the dominant component inthe 802.11 handoff latency and may range from tens to sev-eral hundreds of ms depending on the cards. Note that theMN chooses the AP after the scanning; furthermore, duringthe scanning the MN cannot send nor receive data packets. Inthe following we point out two possible solutions to reducethe packet losses that may occur during the scanning. First,the MN could use two transceivers (as proposed in [19]).One of the transceivers would maintain the connection withoFA while the other scans for better APs. Note that this sce-nario would be similar to the one having all stations using thesame channel. Having only one transceiver, Post-Registrationcould be used similarly to the Smooth Handoff mechanismdescribed in [23]. This mechanism consists of the oFA buffer-ing a certain amountof packets sent to theMN. After the BEThas been established between the nFA and oFA in the Post-

Registration, packets arriving at the oFA during the 802.11handoff destined to the MN could have the chance to reachthe MN along the nFA.

4. Performance evaluation of low latency handoff mechanisms

In this part we analyze and compare the performance of thetwo mechanisms. An analytical model is introduced and usedto compare the delay characteristics and the buffer require-ments for an MN that switches from one access point to an-other. A similar model has been used to investigate otherprotocols (Cellular IP [2], HAWAII [3], Optimized SmoothHandoff [4] and Post-Registration [9]).

4.1. An analytical model for Pre- and Post-Registrationhandoff

In this section we propose a simple analytical model that en-ables us to compute and compare performancecharacteristicsof the Pre- and Post-Registration handoff procedure. In par-ticular, we are interested in the delay distribution of packetsinvolved in a handoff, and in the buffer requirements of therelevant FAs.

Consider the network architecture as depicted in gure 1.We model all routers in the network as ordinary M/M/1



638 BLONDIA ET AL.

queues. Hence, each packet passing through some router hasan exponentially distributed randomservice time, which is as-sumed to both include the processing time in the router andthe transmission time. Moreoverthe response time of a packetis also exponentially distributed.

Now consider an MN movingfromthe oFA to thenFA. Weassumethat a handoff is initiated when theanticipationtriggeroccurs and we denote this instant by t 0. In this section wefurthermore assume that we have a source-triggered handoff,but the model can also be applied to target-triggered handoff.The timing of the relevant L2 triggers (L2-LD and L2-LU) isconsidered to be constant. We dene DLD and DLU as thetime between t 0 and the respective triggers and we have that0 < D LD < D LU.

The performance of the handoff procedures strongly de-pends on the timing of these triggers as will be shown in thenext section. Additionally, there are some important randomtime instants for both methods.

• For Pre-Registration, notably important is the moment theregional registration request arrives at the GFA. From thatmoment on packets will be directed to the nFA insteadof the oFA. In our M/M/1 queuing model this moment(t 1,pre) is random and is distributed as the sum of severalexponential variables (the response times of the routers onthe path from oFA via nFA to GFA), and several constantvalues (the propagation delays on the links between theserouters).

• For Post-Registration, the rst essential random value isthe moment the BET is established. This instant (t 1,post)

is also distributed as the sum of exponential variables (theresponse times of the routers on the path from oFA to nFAand back to oFA) and constants (the propagation delays).A second essential value is the moment the regional reg-istration request arrives at the GFA (t 2,post). Once morethis instant is randomly distributed following the sum of exponential variables (the response times of the routers onthe path from nFA to GFA) and constants (D LU and thepropagation delays).

We now look at a CBR UDP stream originating from a CNwith destination theMN. Suppose that everyT ms a packetar-rives at the GFA (we ignore the jitter introduced by the back-bone network). We can then observe the stream starting froma packet arriving some time before t 0, and compute the dis-tribution of the end-to-end delay (GFA to MN) of this packetand each subsequent packet involved in the handoff.

The delay of each packet is again a random variable madeup from exponential variables and constants, and its spe-cic form depends on the path the packet follows. The lat-ter is in turn dependent on stochastic events involving therandom time instants dened above. Denote by t 1GFA thepredetermined arrival instant of the rst observed packet atthe GFA. The arrival of the kth packet then equals t kGFA =t 1GFA+ (k − 1)T . Denoteby xFA yFA the random time needed

for a packet to travel fromxFA toyFA throughthe connectingrouters. In the case of Pre-Registration the following stochas-

tic events determine the path of the kth packet and thus itsdelay distribution:1. t kGFA < t 1,pre → kth packet routed via the oFA:

(a) t kGFA + GFA oFA < t 0 + D LD → packet forwarded

from oFA to MN,(b) t kGFA + GFA oFA > t 0 + D LD → packet lost,

2. t kGFA > t 1,pre → kth packet routed via the nFA:

(a) t kGFA + GFA nFA < t 0 + D LU → packet needs to bebuffered at nFA, in which case it will be forwarded toMN at t 0 + D LU,

(b) t kGFA + GFA nFA > t 0 + D LU → packet forwardedfrom nFA to MN.

And similarly, for Post-Registration the following schemeholds:1. t kGFA < t 2,post → kth packet routed via the oFA:

(a) t kGFA + GFA oFA < t 0 + D LD → packet forwardedfrom oFA to MN,

(b) t 0 + D LD < t kGFA + GFA oFA < t 1,post → packet

needs to be buffered at oFA, in which case it will beforwarded through the BET at t 1,post,

(c) max(t 0 + D LD, t 1,post) < t kGFA + GFA oFA → packet

forwarded through the BET:

(i) t kGFA+ GFA oFA+ oFA nFA < t 0+ D LU → packetneeds to be buffered at nFA, in which case it willbe forwarded to MN at t 0 + D LU,

(ii) t kGFA+ GFA oFA+ oFA nFA > t 0+ D LU → packetforwarded from nFA to MN,

2. t kGFA > t 2,post → kth packet routedvia the nFA, forwardedfrom nFA to MN.

Taking into account these schemes allows us to computethe overall delay distribution of the kth packet. In particular,denoting by GFAE k

MN the end-to-end delay from GFA to MNfor packet k, we have that

P GFAE kMN > t =

j ∈J

P GFAE kMN > t &(eventj ) ,

where J = { (1a), ( 1b), ( 2a), ( 2b)} for Pre-Registration andJ = { (1a), ( 1b), ( 1ci), ( 1cii), ( 2)} for Post-Registration.

The end-to-end delay takes on different forms accordingto the event it is associated with. As an example, for Post-Registration in case of events (1ci) and (1cii) we have that

(i) GFAE kMN = t 0 + D LU + F MN − t kGFA,

(ii) GFAE kMN = GFA oFA + oFA nFA + F MN,

where F MN denotes the forwarding of the packet from thenFA to the MN, which is modeled as one more response timefor the router nFA. The propagation delay between an FA andthe MN is assumed to be negligible. Based on the M/M/1




assumptions, the delay distribution can now be computed in arather tedious but fairly straightforward way.

Remark that we can compute the delay distribution eitheron the assumption of available buffer size at the FAs or on theassumption of absence of buffers. In the latter case we can,

e.g., compute the expected number of lost packets due to thehandoff (see the next section).To determine the size of the buffer that should be installed

in the FAs in order to avoid or minimize packet loss, we canproceed as follows. We focus on buffering at the nFA, al-thoughfor Post-Registration,a bufferat the oFA could also beconsidered. At the nFA the buffer is needed for packets thatarrive before the LU trigger occurs. (Note that this is assumedto be a separate buffer meant to store the waiting packets des-tined to the MN and this is not modeled as an M/M/1 queue.)The required buffer size can be determined by computing thedistribution of N B, the number of packets that would be lostin the absence of a buffer. If we denote by p

loss(M) the prob-

ability that at least one packet is lost during a handoff at thenFA with buffer capacity M , then we have

p loss(M) = P [N B > M ].

Specic random time intervals I B can be determined forwhich it holds that a packet is lost (in the absence of a wait-ing buffer) if and only if it arrives in that interval at the nFA.These are given by

Pre : I B = [ t 1,pre + GFA nFA; D LU],

Post : I B = max(t 1,post, t 0 + D LD) + oFA nFA; D LU .

RegardingPost-Registrationwe haveassumed here that nowaiting buffer is present at the oFA. Hence the packets thatwould get lost at the nFA are those arriving in the oFA afterthe link is down and after the BET is established, and arrivingvia the BET at the nFA before the link is up.

The respective lengths L B of these intervals have a distri-bution composed of exponential variables and constants, asagain follows from our M/M/1 model. Conditioned on L B itis straightforward to approximate the distribution of N B, sowe can consider

p loss(M) = P [N B > M | L B = z]f L B(z) dz,

where f L B is the density function of L B.The integral can be computed numerically. The latter

equation can then be used as an alternative method to calcu-late the expected number of packets that need buffering andfurthermore we can determine the required buffer size at thenFA by applying

minM : p loss(M) < 10− α .

Here we can choose, e.g., α = 5. In the next section we

present some numerical results concerning M , as well as re-sults for the delay distribution of the stream of packets.

4.2. Numerical results

We assume a network topology as depicted in gure 1. Theresults that will be shown are all obtained with the followingnetwork characteristics. The propagation delays on the links

connecting the GFA and theoFA, as well as the links connect-ing the GFA and the nFA are all set to τ 1 = 5 ms, while thelinks connecting the oFA and the nFA (via router 3) have apropagation delay of τ 2 = 3 ms. The service rate µ in eachrouter is set to 1 packet per ms and all routers have a load of ρ = 0.8. This leads to an exponentially distributed responsetime for each routerwith rate equal toµ( 1− ρ) = 0.2 (ms)− 1.Furthermore we assume that the CN transmits a packet everyT = 10 ms (UDP stream) destined for the MN. These char-acteristics are arbitrarily chosen and are not essential sincethe model is used mainly for comparison between Pre- andPost-Registration.

First we present some results for the delay distribution of a stream of packets involved in a handoff. The start of thehandoff is set to t 0 = 0, and we observe a stream of 30 pack-ets, the rst of which arrives at the GFA at t 1GFA = − 80 ms.We dene the playout time as the maximum allowed end-to-end delay from GFA to MN: if a packet’s end-to-end delayexceeds the playout time it will be dropped.

In gure6 we assume that there is no buffercapacity avail-able, which implies that packets can get lost due to the hand-off. The curves depict the expected number of packets fromthe stream that are dropped due to the expiration of the play-out time or the absence of a buffer. Both the results for Pre-Registration and Post-Registration are shown, for two differ-

ent values of the time between the LD and the LU trigger.The timing of the LD trigger is set to 60 ms since t 0. Notethat these curves tend to the expected number of lost packetsdue to the absence of buffer capacity when the playout timetends to innity. It can be seen that Pre-Registration results inmore losses than Post-Registration, while the average delayfor packets that are not lost is slightly larger for the Post-Registration scheme. The latter follows from the fact that

Figure 6. Delay – Post- and Pre-Registration.



640 BLONDIA ET AL.

Figure 7. Delay – Post-Registration.

Figure 8. Individual delays – Pre.

packets using the BET have a longer delay. When the timebetween the LD and the LU trigger increases, more packetsare lost, so more buffer capacity would be needed to avoidlosses.

In gure 7 we show the impact of providing innite buffersat the FAs. The expected number of dropped or lost packetsin case of innite buffers is compared with the case of theabsence of buffers. For Post-Registration, the packets thatwould be lost when there is no buffer available will obviouslynot be lost when buffers are installed, but they will experi-ence a larger average end-to-end delay. The curve for innitebuffers in case of Pre-Registration does not converge to zerosince packets may get lost in the oFA.

Figures 8 and 9 depict the delay distribution of each indi-vidual packet of the observed stream. In particular the curvespresent the probability that the GFA-MN end-to-end delay of

the kth packet exceeds t . The timing of the LD and LU trig-ger is xed at 80 ms and 120 ms, respectively, and we as-

Figure 9. Individual delays – Post.

sume innite buffer sizes. The curves for the rst few pack-ets and the last few packets are essentially the same, sincethese packets are not inuenced by the handoff. It can thenbe seen that for Post-Registration the inuence of the hand-off on the stream’s delay starts later as well as lasts longer.This corresponds to the fact that Post-Registration(optimally)keeps using the link between the oFA and the MN, while Pre-Registration possibly already redirectspackets before this linkis down. Such packets (e.g., k = 12 and k = 13) then havea longer delay since they have to wait for the LU trigger. Thedisadvantage of Post-Registration is reected by the fact that

it lasts until packet 25 for the inuence of the handoff pro-cedure to have disappeared completely. Note that packets 19to 22 have the same delay distribution. These packets, accord-ing to our model, all travel with a high probability throughthe BET, straight to the MN without having to be buffered.In summary, these gures show that Post-Registration is bet-ter adjusted to the timing of the triggers, but more packetshave a slightly larger delay. A Pre-Registration handoff af-fects fewer packets, but some of these can have unnecessaryhigh delays.

Finally we show some results on the required buffer ca-pacity as a function of the timing of the L2 triggers. Fig-ures 10 and 11 each depict both the expected number of pack-ets that would be lost if there would be no buffer available atthe nFA (thick line), as well as the minimum buffer size M to ensure that P loss(M) < 10− 5 (thin line). Each gure com-pares Pre-Registration (solid line) to Post-Registration (dot-ted line). In gure 10 the timing of the LU trigger is xedat DLU = 200 ms, and DLD is varied. In case of Pre-Registration, only the timing of the LU trigger affects thesituation at the nFA, and not the timing of the LD trigger,hence the horizontal curves. As the time between the LD andLU trigger decreases, the number of packets that need to bebuffered at the nFA in case of Post-Registration decreases aswell. As can be seen from the gure, Post-Registration of-

fers considerable gain over Pre-Registration when the timebetween the triggers is limited.




Figure 10. Buffering in nFA.

Figure 11. Buffering in nFA.

In gure 11 both LD and LU vary, but the time betweenLD and LU is xed. The required buffer capacity for Pre-Registration is obviously linearly related to the timing of the LU trigger, while for Post-Registration it is essentiallythe difference (which is xed here) between the two trig-gers that determines the buffer requirements. ConcerningPost-Registration, the fact that for smaller values of DLU

fewer packets need to be buffered, is explained by notingthat in these cases fewer packets are transmitted through theBET because the BET might not yet be established at thetime of DLD and we did not consider a buffer at the oFA.These last two gures indicate that Pre-Registration in gen-eral, but not necessarily, requires more buffer capacity thanPost-Registration.

5. Scalability of the handoff protocols

5.1. OPNET model description

The OPNET simulation models used in this paper (seealso [26]) are built into OPNET’s standard models [20], ex-

Figure 12. OPNET network model.

tending their functionality and allowing extensive use of the

stock router, link and host models. The implemented pro-tocols are Mobile IP [19] extended with Route Optimiza-tion [22], Hierarchical MIP [13], Optimized Smooth Hand-off [23] and Low Latency Handoffs [10].

Mobility agents are modeled as regular hosts, connectedto routers (indicated as oFA, nFA and RoFA, RnFA, respec-tively, in gure 12), not as one co-located router and accesspoint node. These entities are derived from standard worksta-tion models, but extended with several new processes: mobil-ity registration service, (de-)tunneling service, agent adver-tisement mechanism and proxy ARP (used at the HA). Theseare located on the same IP (sub)network as a Layer 2 bridge,which in this case is an 802.11b WiFi − Ethernet bridge.Mobile nodes are very similar to the agents detailed above asthey are descendant from the same base model and processadditions, but they are equipped with an arbitrary number of Layer 2 interfaces, one for each network the node can roamunto. While they do not need to be of the same type, all inter-faces are 802.11b for the course of this paper.

The handover protocols discussed here rely on a level of integration between IP and the underlying access technologyin a way that the IP layer needs to receive certain input trig-gers from the underlying one. For the sake of evaluating theseprotocols per IETF standards, all triggers are delivered in a“deus ex machina” way. Each MN in the scenario executes

a pre-dened script in which all handoff events described insection 2 are given (either constant time or according to some



642 BLONDIA ET AL.

Figure 13. Peak buffer usage.

distribution). Execution blocks in this script can be repeatedany number of times to allow for very controlled movementscenarios.

Figure 12 shows a typical OPNET setup. Only the hierar-chical branch and the correspondentnode are shown here, notthe home network or any mobile nodes. The foreign agententities are connected to the access points through Ethernetswitches. The GFA is also a simple host on an Ethernetdomain, which is 10 Mbps throughout. Special attention isneeded when conguring this network for a large number of MNs, e.g., ARP cache sizes in the foreign agents must be

large enough to support the number of nodes.5.2. Results for CBR trafc

The CBR trafc consists of constant-sized (500 bytes) UDPpackets sent to all the mobile nodes every 50 ms. The mobil-ity script for each MN is as follows: it starts in one of the twodomains and draws a uniformly distributed random value be-tween 0 and 5 s to initiate its rst move; the MNs are dividedequally among the APs at the start of the simulation. EveryMN then stays a uniformly distributed dwell time (between 4and 6 seconds) in one domain before switching to the other.

In gure 13 the peak buffer usage is measured in bothforeign agents when each MN conducts 100 handovers.Post-Registration needs signicantly less buffers as the timeneeded to buffer packets depends on the spacing of oFAL2-LD and nFA L2-LU since the bi-directional edge tun-nel (BET) starts forwarding trafc the moment the link goesdown. In Pre-Registration however, the anticipation triggerwhich occurs before L2-LD causes the MN to register withits nFA through the oFA. This registration goes up to the GFAwho will change the routing path for the MN’s trafc.

Figures 14 and 15 show peak buffer usage measured in theforeignagents with a constant number of MNs butwhen vary-ing both the nFA L2-LU and the oFA L2-LD trigger. In g-

ure 14 the anticipation trigger arrives at 0 s, and oFA L2-LDprecedes the LU event by 100 ms. In gure 15, oFA L2-LD

Figure 14. Peak buffer usage varying L2-LU.

Figure 15. Peak buffer usage varying L2-LD.

varies between the anticipation trigger and nFA L2-LU at500 ms. They clearly illustrate the dependence of buffer re-quirements on one of the triggers: Post-Registration is notinuenced by the timing of the anticipation trigger, as long asthere is time enoughforBET setup (this is dictated by the pro-tocol requirements). Similarly, Pre-Registration’s buffer re-quirements remain unaffected when varying the oFA L2-LDtrigger. Remark that these results are completely in corre-spondence with the conclusions drawn from the analytical re-sults (see gures 10 and 11).

A certain amount of trafc delay is of course associatedwith these handovers. Apart from the delay experienced dueto buffering, which will usually be higher in Pre-Registration,the Post-Registration will have an additional delay due to useof the BET. This tunnel is used until the MN node choosesto register. Neither Pre- nor Post-Registration induce an extra

delay when the number of nodes is increased: the buffers inplace are logical per-node buffers.




Figure 16. TCP goodput.

5.3. Results for TCP trafc

TCP trafc consists of FTP le transfers (10 MB) to eachof the mobile nodes. FTP response time is then measuredat each of the mobile nodes (time between start and end of the le request) and averaged over the total of nodes. TCPReno is used in all the runs, but the MSS has been reduced to1440 bytes to avoid fragmentation induced by tunneling (anextra IP header is added to tunneled packets). Ethernet linksare loaded with background trafc to avoid overloading thewireless channel.

Figure 16 shows the advantage of using a low latency han-dover method as opposed to plain MIP. TCP ows clearlybenet from smoothness of the handoffs, but there is no sig-nicant difference between the two methods discussed here.Post-Registration might behave slightly worse because thereis a possibility of packet re-sequencing: packets following theold route through the tunnel might arrive later than packetssent directly to the nFA if the MN has just performed nor-mal registration. Hence, the advantage of Post-Registrationwith respect to packet loss is reduced by the requirement of re-sequencing.

The inuence of trigger timing on TCP’s performance isnegligible: the time that packets spend in the FA’s buffer is notlarge enough to cause performance degradation as the time-out values are larger: TCP commonly has an absolute bottomvalue of 500 ms for timeout timers. This is illustrated in g-ures 17 and 18.

5.4. Impact of network topology

Let us now consider a tree-structured network depicted in g-ure 19 instead of the reference network used throughout thispaper. This shows a more tree-like physical setup and is iden-tical to the architecture of gure 12, except for R3 which hasbeen moved up one level to become the crossover router CR.

Using the mobility scenario described above, we can ex-amine protocol performance differences between these two

Figure 17. TCP goodput, Pre-Reg.

Figure 18. TCP goodput, Post-Reg.

networks for, e.g., Pre-Registration. To emphasize the im-pact of the router R3, the 1-hop distance (delay) between oFAand nFA in the reference network has been made very small.Figure 20 shows that the original network has higher bufferrequirements: when the routing distance between oFA andnFA is very small, the registration request to the nFA routed

through the oFA will reach the GFA much sooner, resultingin more packets being forwarded for the same trigger values.However, both curves behave very similarly.

6. Conclusion

In this paper, we have analyzed two Layer 3 low latency mo-bility protocols, Pre- and Post-Registration. Their operationrelies on the availability of L2 triggers. Possible implemen-tations when using IEEE 802.11 as link layer are discussedand illustrated. A simple analytical model is presented thatallows a detailed investigationof the delay characteristics andthe buffer requirements for a single mobile node involved ina handoff. Although the analytical and simulation results



644 BLONDIA ET AL.

Figure 19. Alternate network topology.

Figure 20. Peak buffer usage for both architectures.

show that the two schemes perform basically the same, wenoted some differences which can be explained intuitively.Pre-Registration needs in general more buffers to obtain zero

losses and moreover, packets may get lost in the oFA. Withrespect to delay, Post-Registration is better adjusted to the

timing of the triggers, but more packets have a slightly largerdelay. A Pre-Registration handoff affects fewer packets, butsome of these can have unnecessary high delays. OPNETsimulations conrm these conclusions in more realistic sys-tems with respect to the number of mobile nodes and their

mobility pattern. The difference in TCP throughput for bothschemes is negligible.

References

[1] ANSI/IEEE Std 802.11 (1999 ed.).[2] C. Blondia, O. Casals, P. De Cleyn and G. Willems, Performance analy-

sis of IP micro-mobility handoff protocols, in: Proceedings of Pro-tocols for High Speed Networks 2002 (PfHSN 2002) , Berlin (2002)pp. 211–226.

[3] C. Blondia, O. Casals, Ll. Cerdà and G. Willems, Performance analysisof a forwarding scheme for handoff in HAWAII, in: Proceedings of Networking 2002 , Pisa (2002) pp. 504–514.

[4] C. Blondia, O. Casals, N. Van den Wijngaert and G. Willems, Perfor-mance analysis of smooth handoff in Mobile IP, in: Proceedings of theFifth ACM International Workshop on Modeling, Analysis and Simu-lation of Wireless and Mobile Systems (MSWiM2002) , Atlanta, USA(September 2002).

[5] C. Blondia, O. Casals, Ll. Cerdà, N. Van den Wijngaert, G. Willemsand P. De Cleyn, Low latency handoff mechanisms and their imple-mentation in an IEEE 802.11 network, in: Proceedings of ITC (2003)(to appear).

[6] A.T. Campbell, J. Gomez, C.-Y. Wan, S. Kim, Z. Turanyi and A. Valko,Cellular IP, draft-ietf-mobileip-cellular-00-txt (December 1999).

[7] A.T. Campbell, J. Gomez, C.-Y. Wan, S. Kim, Z. Turanyi and A. Valko,Cellular IP performance, draft-gomez-cellularip-perf-00.txt (October1999).

[8] A.T. Campbell and J. Gomez-Castellanos, IP micro mobility protocols,ACM SIGMOBILE Mobile Computing and Communications Review4(4) (October 2000) 45–54.

[9] O. Casals, Ll. Cerdà, G. Willems, C. Blondia and N. Van den Wijngaert,Performance evaluation of the Post-Registration method, a low latencyhandoff in MIPv4, in: Proceedings of ICC (2003) (to appear).

[10] K. El Malki et al., Low latency handoffs in mobile IPv4, IETF draft-ietf-monileip-lowlatency-handoffs-v4-04.txt (2002).

[11] K. El Malki and H. Soliman, Simultaneous bindings for mobile IPv6fast handovers, IETF draft-elmalki-mobileip-bicasting-v6-03.txt (May2003).

[12] S. Goswami, Simultaneous handoff of Mobile-IPv4 and 802.11, IETFdraft-goswami-mobileip-simultaneous-handoff-v4-01.txt (2002).

[13] E. Gustafsson, A. Jonsson and C. Perkins, Mobile IP regional registra-tion, draft-ietf-mobileip-reg-tunnel-07.txt (October 2002).

[14] H. Haverinen and J. Malinen, Mobile IP regional paging, draft-haverinen-mobileip-reg-paging-00.txt (June 2000).

[15] J. Kempf, P.R. Calhoun and C. Pairla, Foreign agent assisted handoff,draft-calhoun-mobileip-proactive-fa-01.txt (June 2000).

[16] P. McCann, Mobile IPv6 fast handovers for 802.11 networks, IETFdraft-mccann-mobileip-80211fh-00.txt (2002).

[17] A. Mishra, M. Shin and W. Arbaugh, An empirical analysis of the IEEE802.11 MAC layer handoff process, Technical Report CS-TR-4395,Department of Computer Science, University of Maryland (September2002).

[18] ns with micromobility support, http://comet.ctr.columbia.edu/ micromobility

[19] M. Ohta, Smooth handover over IEEE 802.11 wireless LAN, draft-ohta-smooth-handover-wlan-00.txt (June 2002).

[20] OPNET Modeler, www.opnet.com[21] C.E. Perkins, IP mobility support, rfc 2002 (October 1996).[22] C. Perkins and D. Johnson, Route optimization in mobile IP, IETF

draft-ietf-mobileip-optim-11.txt (September 2001).




[23] C. Perkins and K.-Y. Wang, Optimized smooth handoffs in mobile IP,in: Proceedings of IEEE Symposium on Computers and Communica-tions , Egypt (July 1999).

[24] R. Ramjee, T. La Porta, S. Thuel, K. Varadhan and L. Salgarelli, IPmicro-mobility support using HAWAII, draft-ietf-mobileip-hawaii-00.txt (June 1999).

[25] The network simulator – ns-2, http://www.isi.edu/nsnam/ns[26] N. Van den Wijngaert and C. Blondia, Integration of IP mobility inOPNET: Modeling and simulation, in: Proceedings of OPNETWORK 2002 , Washington, DC (August 2002).

[27] F. Vena, Ll. Cerdà and O. Casals, Study of the TCP dynamics overwireless networks with micromobility support using the ns simulator,in: European Wireless 2002 , Florence (February 2002) pp. 114–120.

[28] A.E. Yegin et al., Supporting optimized handover for IP mobility– requirements for underlying systems, IETF draft-manyfolks-12-mobilereq-02.txt (2002).

Chris Blondia obtained his Ph.D. in mathematicsfrom the University of Ghent (Belgium) in 1982.Between 1986 and 1991 he was a researcher at thePhilips Research Laboratory and from 1991 till 1994he was an Associate Professor at the Computer Sci-ence Department of the University of Nijmegen (TheNetherlands). In 1995 he joined the Department of Mathematics and Computer Science of the Univer-sity of Antwerp, where he is currently a Professor,head of the research group “Performance Analysis

of Telecommunication Systems” (PATS) and head of the Department. Hismain research interests are related to both theoretical methods for stochasticmodelling (in particular queueing systems) and to the design, the modellingand performance evaluation of telecommunication systems, in particular re-lated to trafc management in broadband networks, IP mobility management,medium access control for wireless and wired access networks, etc. He haspublished a substantial number of papers in international journals and con-ference proceedings. He is editor of the “Journal of Network and ComputerApplications” and member of IFIP Working Group 6.3 on “Performance of Computer Networks”. He has been involved in several European researchprograms (RACE, ACTS, IST and COST).E-mail: [email protected]

Olga Casals obtained the engineering degree intelecommunication in 1983 from the PolytechnicUniversity of Catalonia, and received the Ph.D. de-gree from the same University in 1986. Since the endof 1983she hasbeen working at the Computer Archi-tecture Department of the Polytechnic University of Catalonia where she is head of a research group fortrafc research in B-ISDN communication systems.In 1994 she became Full Professor. From May 1998till March 2001 she was Head of the Department of

Computer Architecture. Her areas of interest include computer and commu-nication system performance evaluation and B-ISDN trafc control mecha-nisms. Olga Casals is a member of IFIP Working Group 6.3 on Performanceof Communication Systems. She has published a substantial number of pa-pers in international journals and conferences on performance evaluation of computer and communication systems and on trafc control in B-ISDN net-

works. She has participated in many projects (RACE, ACTS, IST, COST)funded by the European Commission.E-mail: [email protected]

Llorenç Cerdà obtained the engineering degree intelecommunications in 1993 from Polytechnic Uni-versity of Catalonia (UPC). He joined the ComputerArchitecture Department of UPC in 1994, where re-ceived the Ph.D. degree in 2000 and currently heis Assistant Professor. His areas of interest includeTCP, IP micromobility, wireless 802.11 networks,C++ programming, simulation, performance eval-uation, queueing networks. He has participated inseveral EU funded research projects (ACTS, IST,

COST), research projects in collaboration with Nokia (TESI), and the startupSouthWing.

Nik Van den Wijngaert obtained his degree inmathematics and computer science from the Univer-sity of Antwerp, Belgium in 2001. Since then hehas been doing research as a Ph.D. student in thePATS (Performance Analysis of TelecommunicationSystems) group, which is part of the Department of Mathematics and Computer Science. His research

interests include simulation, IP mobility and wirelesscommunications.

Gert Willems graduated in mathematics from theUniversity of Antwerp, Belgium, in 2000. He is cur-rently a research assistant and Ph.D. student in theDepartment of Mathematics and Computer Scienceat the University of Antwerp. His research interestsinclude modeling and simulation of communicationnetworks and computational and robust statistics.

layer 3 low latency handoff

Documents